def __init__(self, cfg):
super(Encoder, self).__init__()
self.cfg = cfg
# Layer Definition
vgg16_bn = vgg16(pretrained=False, batch_norm=True)
self.vgg = paddle.nn.Sequential(
*list(vgg16_bn.features.children())[:27])
self.layer1 = paddle.nn.Sequential(
paddle.nn.Conv2D(
512,
512,
kernel_size=1,
weight_attr=paddle.nn.initializer.KaimingNormal(),
bias_attr=paddle.nn.initializer.Constant(value=0.0)),
paddle.nn.BatchNorm2D(512),
paddle.nn.ELU(),
)
self.layer2 = paddle.nn.Sequential(
paddle.nn.Conv2D(
512,
256,
kernel_size=3,
weight_attr=paddle.nn.initializer.KaimingNormal(),
bias_attr=paddle.nn.initializer.Constant(value=0.0)),
paddle.nn.BatchNorm2D(256), paddle.nn.ELU(),
paddle.nn.MaxPool2D(kernel_size=4))
self.layer3 = paddle.nn.Sequential(
paddle.nn.Conv2D(
256,
128,
kernel_size=3,
weight_attr=paddle.nn.initializer.KaimingNormal(),
bias_attr=paddle.nn.initializer.Constant(value=0.0)),
paddle.nn.BatchNorm2D(128), paddle.nn.ELU())
def __init__(self,
generator,
discriminator=None,
cycle_criterion=None,
idt_criterion=None,
gan_criterion=None,
l1_criterion=None,
l2_criterion=None,
pool_size=50,
direction='a2b',
lambda_a=10.,
lambda_b=10.,
is_train=True):
"""Initialize the PSGAN class.
Parameters:
cfg (dict)-- config of model.
"""
super(MakeupModel, self).__init__()
self.lambda_a = lambda_a
self.lambda_b = lambda_b
self.is_train = is_train
# define networks (both Generators and discriminators)
# The naming is different from those used in the paper.
# Code (vs. paper): G_A (G), G_B (F), D_A (D_Y), D_B (D_X)
self.nets['netG'] = build_generator(generator)
init_weights(self.nets['netG'], init_type='xavier', init_gain=1.0)
if self.is_train: # define discriminators
vgg = vgg16(pretrained=False)
self.vgg = vgg.features
cur_path = os.path.abspath(os.path.dirname(__file__))
vgg_weight_path = get_path_from_url(VGGFACE_WEIGHT_URL, cur_path)
param = paddle.load(vgg_weight_path)
vgg.load_dict(param)
self.nets['netD_A'] = build_discriminator(discriminator)
self.nets['netD_B'] = build_discriminator(discriminator)
init_weights(self.nets['netD_A'],
init_type='xavier',
init_gain=1.0)
init_weights(self.nets['netD_B'],
init_type='xavier',
init_gain=1.0)
# create image buffer to store previously generated images
self.fake_A_pool = ImagePool(pool_size)
self.fake_B_pool = ImagePool(pool_size)
# define loss functions
if gan_criterion:
self.gan_criterion = build_criterion(gan_criterion)
if cycle_criterion:
self.cycle_criterion = build_criterion(cycle_criterion)
if idt_criterion:
self.idt_criterion = build_criterion(idt_criterion)
if l1_criterion:
self.l1_criterion = build_criterion(l1_criterion)
if l2_criterion:
self.l2_criterion = build_criterion(l2_criterion)
def __init__(self, cfg):
"""Initialize the PSGAN class.
Parameters:
cfg (dict)-- config of model.
"""
super(MakeupModel, self).__init__(cfg)
# define networks (both Generators and discriminators)
# The naming is different from those used in the paper.
# Code (vs. paper): G_A (G), G_B (F), D_A (D_Y), D_B (D_X)
self.nets['netG'] = build_generator(cfg.model.generator)
init_weights(self.nets['netG'], init_type='xavier', init_gain=1.0)
if self.is_train: # define discriminators
vgg = vgg16(pretrained=False)
self.vgg = vgg.features
cur_path = os.path.abspath(os.path.dirname(__file__))
vgg_weight_path = get_path_from_url(VGGFACE_WEIGHT_URL, cur_path)
param = paddle.load(vgg_weight_path)
vgg.load_dict(param)
self.nets['netD_A'] = build_discriminator(cfg.model.discriminator)
self.nets['netD_B'] = build_discriminator(cfg.model.discriminator)
init_weights(self.nets['netD_A'],
init_type='xavier',
init_gain=1.0)
init_weights(self.nets['netD_B'],
init_type='xavier',
init_gain=1.0)
self.fake_A_pool = ImagePool(
cfg.dataset.train.pool_size
) # create image buffer to store previously generated images
self.fake_B_pool = ImagePool(
cfg.dataset.train.pool_size
) # create image buffer to store previously generated images
# define loss functions
self.criterionGAN = GANLoss(
cfg.model.gan_mode) #.to(self.device) # define GAN loss.
self.criterionCycle = paddle.nn.L1Loss()
self.criterionIdt = paddle.nn.L1Loss()
self.criterionL1 = paddle.nn.L1Loss()
self.criterionL2 = paddle.nn.MSELoss()
self.build_lr_scheduler()
self.optimizers['optimizer_G'] = build_optimizer(
cfg.optimizer,
self.lr_scheduler,
parameter_list=self.nets['netG'].parameters())
self.optimizers['optimizer_DA'] = build_optimizer(
cfg.optimizer,
self.lr_scheduler,
parameter_list=self.nets['netD_A'].parameters())
self.optimizers['optimizer_DB'] = build_optimizer(
cfg.optimizer,
self.lr_scheduler,
parameter_list=self.nets['netD_B'].parameters())
def __init__(self, in_channels=3):
super(Vgg16Base, self).__init__()
features = vgg16(pretrained=True).sublayers()[0].sublayers()
if in_channels != 3:
features[0] = nn.Conv2D(in_channels,
64,
kernel_size=[3, 3],
padding=1,
data_format='NCHW')
self.features = nn.LayerList(features)
self.features.eval()
def __init__(self):
super(StyleTransferModel, self).__init__()
# pretrained设置为true，会自动下载imagenet上的预训练权重并加载
vgg = vgg16(pretrained=True)
self.base_model = vgg.features
for p in self.base_model.parameters():
p.stop_gradient = True
self.layers = {
'0': 'conv1_1',
'5': 'conv2_1',
'10': 'conv3_1',
'17': 'conv4_1',
'19': 'conv4_2', ## content representation
'24': 'conv5_1'
}